Journal of Economic PerspectivesVolume 31, Number 2Spring 2017Pages 5986

Twenty Years of Time Series

Econometrics in Ten Pictures

James H. Stock and Mark W. Watson

T
wenty years ago, empirical macroeconomists shared some common under-
standings. One was that a dynamic causal effectfor example, the effect on
output growth of the Federal Reserve increasing the federal funds rateis
properly conceived as the effect of a shock, that is, of an unanticipated autonomous
change linked to a specific source. Following Sims (1980), the use of vector autore-
gressions to estimate the dynamic causal effect of shocks on economic variables was
widespread. There was also an understanding that vector autoregressions, because
they impose as little structure on the data as possible, cannot answer questions about
changes in policy regimes, such as the macroeconomic consequences of the Fed
adopting a new policy rule. For such questions, more structured models grounded
in economic theory are needed. At the same time, there was an increasing recog-
nition that the available methods needed significant work. The schemes used to
identify structural shocks in vector autoregressions were often seen as unconvincing
by researchers outside the field, and the small structural models of the time were not
econometrically estimated, miring that enterprise in an unhelpful debate over how
to calibrate such models. In addition, there were chinks emerging in the theoretical
econometric underpinnings of inference in time series data, as well as opportunities
for using the much larger datasets becoming available, if only the tools to do so
could be developed. The time was ripe for progress.

James H. Stockis the Harold Hitchings Burbank Professor of Political Economy,Harvard

University, Cambridge, Massachusetts. Mark W. Watson is Howard Harrison and Gabrielle
Snyder Beck Professor of Economics and Public Affairs, Princeton University, Princeton, New
Jersey. Their email addresses are james_stock@harvard.edu and mwatson@princeton.edu.

For supplementary materials such as appendices, datasets, and author disclosure statements, see the
article page at
https://doi.org/10.1257/jep.31.2.59 doi=10.1257/jep.31.2.59
60 Journal of Economic Perspectives

This review tells the story of the past 20 years of time series econometrics
through ten pictures. These pictures illustrate six broad areas of progress in time
series econometrics: estimation of dynamic causal effects; estimation of dynamic
structural models with optimizing agents (specifically, dynamic stochastic equilib-
rium models); methods for exploiting information in big data that are specialized
to economic time series; improved methods for forecasting and for monitoring
the economy; tools for modeling time variation in economic relationships; and
improved methods for statistical inference.
These pictures remind us that time series methods remain essential for shoul-
dering real-world responsibilities. The world of business, finance, and government
needs reliable information on where the economy is and where it is headed. Policy-
makers need analysis of possible policies, and macroeconomists need to improve
their understanding of the workings of modern, evolving economies. Taken
together, the pictures show how 20 years of research have improved our ability to
undertake these professional responsibilities. These pictures also remind us of the
close connection between econometric theory and the empirical problems that
motivate the theory, and of how the best econometric theory tends to arise from
practical empirical problems.
A review of 20 years of research must make some arbitrary decisions. One of our
decisions is to focus on empirical macroeconomics, not finance. Fortunately, there
are good surveys of the many developments in financial econometrics: for example,
see the papers in At-Sahalia and Hansen (2010). Another concerns the choice of
figures. Our ten figures are not meant to single out superstar papers (although
some are) but rather to represent important lines of research: each figure illus-
trates a broader research program. In choosing these figures, we first looked for
influential early papers from the late 1990s and early 2000s that framed subsequent
research. This yielded five figures from papers with an average of 1,486 Google
Scholar citations each. We then looked for figures more recently published that
illustrate key findings or methods in a relatively mature line of research, yielding
four more figures. Our final figure, which is not from published research, illustrates
an open empirical challenge for research ahead.

Causal Inference and Structural Vector Autoregressions

An ongoing question in empirical macroeconomics is how to determine the

causal effect of a policy change. For example, what is the effect of an autonomous,
unexpected, policy-induced change in the monetary policy target ratethat is, a
monetary policy shockon output, prices, and other macro variables? The under-
lying problem is simultaneous causality: for example, the federal funds interest rate
depends on changes in real GDP through a monetary policy rule (formal or informal),
and GDP depends on the federal funds interest rate through induced changes in
investment, consumption, and other variables. Thus, one cannot determine the effect
of a change in the federal funds interest rate simply by using the rate (perhaps along
with lagged values of the rate) as a right-hand-side variable in a regression to explain
 James H. Stock and Mark W. Watson 61

GDP. Somehow, a researcher needs to isolate the exogenous variation in the federal
funds interest rate, and for that you need external information.
Since the seminal work of Sims (1980), vector autoregressions have been
a standard tool for estimating the causal effects over time of a shock on a given
macro variable. This tool evolved into structural vector autoregressions, which
are based on the idea that the unanticipated movements in the variablesthat is,
their forecast errorsare induced by structural shocks. The goal of structural vector
autoregressions is to impose sufficient restrictions so that one or more structural
shocks can be identified: specifically, that one or more shocks can be represented
as an estimable linear combination of the forecast errors. The result of this analysis
is the estimation of a dynamic path of causal effects, which in macroeconometrics is
called a structural impulse response function.
However, many applications of the original methods for identification of struc-
tural autoregressions that were dominant in the 1980s and 1990s have not withstood
close scrutiny (as articulated, for example, by Rudebusch 1998). For example, a
popular method for identifying monetary policy shocks in the 1980s and 1990s was
to assume that economic activity and prices respond to a monetary policy shock
with a lag, but that monetary policy responds systematically to contemporaneous
nonmonetary shocks to the other variables. Under this assumption, the predicted
value in a regression of the federal funds rate on its lags and on current and lagged
values of the other variables is the endogenous policy response, and the residual is
the unanticipated exogenous componentthat is, the monetary policy shock. But
this identifying assumption is not credible if the other variables include other asset
prices, such as long-term interest rates.
Thus, this area needed new approaches. Broadly speaking, these new
approaches bring to bear external information: information outside the linear
system of equations that constitutes the vector autoregression. The development
of new methods for estimating causal effects has been one of the main advances in
microeconometrics over the past two decades (as discussed in several other articles
in this symposium), and the focus on credible identification has parallels in the
structural vector autoregression literature.

Using External Information to Estimate the Shock Directly

This brings us to our first picture, which is from Kuttner (2001). Kuttners
interest was in estimating the dynamic causal effect of a monetary policy shock on
long-term interest rates, which is part of the broader program of estimating their
dynamic causal effect on macroeconomic variables. Because the Fed controls the
federal funds interest rate, one might initially think that the fed funds rate is exoge-
nous; but not so, because some of the changes are responses to changes in economic
activity which have their own effect on long-term interest rates. Rather, the exog-
enous part of the fed funds ratethe monetary policy shockis the part that is not
a response to economic activity. Kuttners innovation was to draw on external infor-
mation to identify the shock. Specifically, he knew that the Federal Reserve Open
Market Committee announced its decisions at a specific time after its meetings, and
he also had evidence (along with the theory of efficient financial markets) that the
62 Journal of Economic Perspectives

Figure 1
Changes in the 5-year Treasury Rate and in the Target Federal Funds Rate on
Federal Reserve Open Market Committee (FOMC) Announcement Dates
(fed funds changes are decomposed into anticipated and unanticipated components using
changes in the fed funds futures market on announcement dates)

Source: Kuttner (2001), Figure 2.

Note: The figure shows changes in the 5-year Treasury rate (on the y-axes) and in the fed funds target
rate (on the x-axes) on Federal Reserve Open Market Committee (FOMC) announcement dates.
Unanticipated changes in the fed funds ratewhich are the monetary policy shocksare identified as
changes in the fed funds futures rate from before to after the announcement of a change in the FOMCs
target for the fed funds rate. The anticipated change is the actual change in the fed funds futures rate,
minus the unanticipated change.

fed funds future rate was an efficient forecast of future fed funds rates. Thus, he was
able to measure the unexpected part of the change in the federal funds futures rate
as the change in the fed funds rate before and after the announcement. Assuming
that no other relevant news was released during the announcement window, this
change in the fed funds futures rate measures the change in market expectations
of the fed funds rate resulting from the announcementthat is, it measures the
monetary policy shock associated with the announcement. By using this external
information, he could directly estimate the monetary policy shock.
Kuttners figure (our Figure 1) shows that this unanticipated component of
the change in the target rate is associated with changes in the five-year Treasury
rate (right panel), but anticipated changes are not (center). As a result, there is no
particular relationship between the actual announced target and the five-year rate
(left). We interpret this figure as a compelling plot of the first stage in instrumental
variables regression: it shows that an instrument (the unanticipated component of
the target change on the announcement day) is correlated with an endogenous vari-
able (the five-year interest rate).
The idea of using external information to identify shocks for structural vector
autoregression analysis traces back to Romer and Romer (1989), who used textual
and historical information to identify some exogenous monetary policy shocks.
In addition to Kuttner (2001), Cochrane and Piazzesi (2002), and Faust, Rogers,
 Twenty Years of Time Series Econometrics in Ten Pictures 63

Swanson, and Wright (2003), and Bernanke and Kuttner (2005) are early papers
that use interest rate changes around Federal Reserve announcement dates to iden-
tify monetary policy shocks. In a similar spirit, Hamilton (2003) and Kilian (2008)
use external information on international oil supply disruptions to estimate the
effect of oil supply shocks on the economy.
This line of attack aims to measure the exogenous shock directly from
external information, such as knowledge of the interest rate markets around
announcement dates. If the shock can actually be measured, then estimation of
structural impulse response functions is straightforward: because the shock is
uncorrelated with other shocks, one can simply regress a variable of interest on
current and lagged values of the shock, and the resulting coefficients trace out the
dynamic causal effect (for example, Stock and Watson 2011, chap. 15). But doing
so requires a particular strong form of external information: that the shock can
be accurately measured.

Identification by External Instruments

If the external information succeeds in measuring only part of the shock
or produces a noisy measurement of the shock, then the measured shock has the
interpretation as an instrumental variable and regressions on the measure have the
interpretation as the first stage in two-stage least squares. Arguably, many of the shock
measures proposed to date yield imperfect measures. For example, changes in federal
funds futures around an announcement reveal only a part of the monetary policy
shock. In this case, the external shock measure is an instrumental variable: it is exog-
enous (that is, it is uncorrelated with other structural shocks) if properly constructed,
and it is relevant because it is correlated with the true shock. Hamilton (2003) uses his
measured international oil shock measure as instrument in a single-equation setting.
In a vector autoregression, the technicalities differ from standard instrumental vari-
ables regression because the observed endogenous variables are forecast errors, not
the original variables themselves. Still, the two criteria for a valid instrument, rele-
vance and exogeneity, are the same in the structural vector autoregression application
as in standard instrumental variable regression.
The explicit use of external instruments in structural vector autoregressions is
fairly recent. This method is described in Stock (2008), Ramey (2016), and Stock
and Watson (2016). Empirical applications of identification of structural impulse
response functions using external instruments include Stock and Watson (2012a),
Mertens and Ravn (2013), and Gertler and Karadi (2015).

Identification by Heteroskedasticity
Another method for identifying impulse response functions developed during
the past 20 years exploits the observation that changes in the variance of the shocks
can serve to identify the impulse response functions if those responses remain
constant despite the heteroskedasticity of the shocks. Suppose that there are two
known regimes, a high- and a low-volatility regime. Identification by heteroskedasticity
works by generating two sets of moment equations, one for each regime. Although
64 Journal of Economic Perspectives

neither set can be solved on its own (the identification problem), assuming that the
impulse response functions are the same across both regimes imposes enough para-
metric restrictions that together the two sets of equations can be solved, and thus
the impulse response functions can be identified. This clever insight was developed
for regime-shift heteroskedasticity by Rigobon (2003) and Rigobon and Sack (2003,
2004), and for conditional heteroskedasticity by Sentana and Fiorentini (2001) and
Normandin and Phaneuf (2004). Ltkepohl (2013) offers a survey and discussion.

Identification by Sign Restrictions

An altogether different approach to identification in structural vector autore-
gressions is to use restrictions on the sign of impulse responses to identify the
economic shocks. For many shocks, disparate macro theories often agree on the
signs of their effects, at least over short horizons. Although several early papers
build off this insight, the method developed by Uhlig (2005) is the most widely
used. In his application, Uhlig restricted the impulse response with respect to a
monetary policy shock identified by requiring that, on impact and over the next
five months, the response of overall prices, commodity prices, and nonborrowed
reserves to a contractionary monetary policy shock are not positive, and that the
response of the federal funds interest rate is not negative. Identification using sign
restrictions can be compelling and has been widely adopted.
At a mathematical level, using sign restrictions is fundamentally different than
the other methods that identify shocks: with enough restrictions, those methods lead,
in large samples, to a unique impulse response function, whereas the sign restric-
tions approach only determines a set that includes the impulse response. That is,
sign-identified impulse response functions are not point-identified, but instead are
set-identified.
Set identification of impulse response functions raises subtle issues of inference,
which have only recently been appreciated. Following Uhlig (2005), the standard
approach is Bayesian, but just as the identification scheme in classical structural
vector autoregression methods can strongly influence results, the prior distribution
over the unidentified region of the impulse response parameter space strongly influ-
ences Bayesian inference, even in large samples. These methods therefore require
great care to produce transparent, valid, and robust inference. Recent papers tack-
ling inference in sign-identified structural vector autoregressions are Fry and Pagan
(2011), Moon, Schorfheide, and Granziera (2013), Giacomini and Kitagawa (2014),
Baumeister and Hamilton (2015), and Plagborg-Mller (2016). For additional discus-
sion and references to the recent methodological literature see Stock and Watson
(2016, Section 4).

Estimation of Dynamic Stochastic General Equilibrium Models

Dynamic stochastic general equilibrium models are models of forward-looking,

optimizing economic agents who live in an economy subject to unexpected shocks.
The development of methods for solving and estimating these models, combined
 James H. Stock and Mark W. Watson 65

with their grounding in optimizing economic theory, has made them a central tool
of monetary policy analysis at central banks.
One of the first full-system estimations of a dynamic stochastic general equilib-
rium model was by Ireland (1997), who estimated a three-equation (GDP, prices, and
money) system by maximum likelihood. However, maximizing the likelihood proves
far more difficult numerically than averaging over the likelihood using a Bayesian
prior, and today the dominant methods for estimating dynamic stochastic general
equilibrium models are Bayesian. These methods were first used by DeJong, Ingram,
and Whiteman (2000), Schorfheide (2000), and Otrok (2001) for small dynamic
stochastic general equilibrium systems. Smets and Wouters (2003) showed that these
methods can be applied to larger dynamic stochastic general equilibrium models that
are rich enough to be a starting point for monetary policy analysis.
Figure 2, taken from Smets and Wouters (2003), represents the breakthroughs
made over the past 20 years in the estimation of dynamic stochastic general equi-
librium models. In their model, the Calvo wage parameter in the first panel is
the probability that a workers wage does not change, and the Calvo price param-
eter in the second panel is the probability that the firms price does not change.
As Figure 2 illustrates, the method works: The computational problems encoun-
tered when fitting dynamic stochastic general equilibrium models using frequentist
methods such as maximum likelihood are sidestepped by computing posteriors,
facilitated by a suite of tools developed in the modern Bayesian computational liter-
ature. For some parameters, such as the Calvo price parameter, the data are highly
informative: incorporating the data results in much stickier prices than the authors
prior, so that the posterior and prior distributions are quite different. But for other
parameters, such as the Calvo wage parameter, the data are much less informative,
so that the prior and posterior essentially coincide. Thus, the Calvo wage parameter
is in effect calibrated by the researcher, so the resulting complete model combines
estimation where the data are informative with calibration where they are not.
This property of estimation cum calibration means that care needs to be taken
in interpreting measures of uncertainty arising from the model. From a frequentist
perspective, a classic justification of Bayesian methods is that coverage intervals (Bayes
credible sets) computed using the Bayesian posterior are essentially the same as
frequentist confidence intervals in large samples, as long as a continuous prior does not
rule out parameter values. (This is the celebrated Bernsteinvon Mises theorem.) But
for dynamic stochastic general equilibrium models, because the data are uninformative
for some parametersthat is, some parameters are poorly identifiedthis equivalence
does not hold and the uncertainty measures are heavily influenced by the shape of the
prior. We return to this issue below, when we discuss weak identification.
The literature on estimation of dynamic stochastic general equilibrium models
is vast and, because it quickly gets into specialized computational devices, it can be
difficult to penetrate. For example, models of the SmetsWouters sort rely on log-
linearized approximations to decision rules, which both makes the models fairly
easy to solve and means that the Kalman filter can be used to compute the Gaussian
likelihood. Much of the recent methodological research on estimation of these
models has focused on avoiding the log-linearization step. Among other things,
66 Journal of Economic Perspectives

Figure 2
Prior and Posterior Distributions for Two Structural Parameters in a Dynamic
Stochastic General Equilibrium Model

Source: Smets-Wouters (2003), Figure 1c (upper panel).

Note: This figures represents the breakthroughs made over the past 20 years in the estimation of dynamic
stochastic general equilibrium models. In the model of Smets-Wouters (2003), the Calvo wage
parameter in the first panel is the probability that a workers wage does not change, and the Calvo price
parameter in the second panel is the probability that the firms price does not change.

avoiding log-linearization can improve the ability to analyze the effects of risk
and uncertainty. However, there are substantial computational challenges in esti-
mating nonlinear models, so that log-linearization remains common in practice.
Canova (2007) provides an accessible textbook treatment of the linearize/Kalman
filter/Bayes approach. Herbst and Schorfheide (2015) provide an up-to-date
textbook treatment that focuses on computationally efficient methods for evalu-
ating the posterior of linearized models. Fernndez-Villaverde, Rubio-Ramrez,
and Schorfheide (2016) provide a detailed overview of methods that avoid
linearization.
 Twenty Years of Time Series Econometrics in Ten Pictures 67

Dynamic Factor Models and Big Data

The idea of using a large number of series to understand macroeconomic

fluctuations is an old one, dating back at least as far as the economic indexes and
forecasts of the Harvard Economic Service in the 1920s (Friedman 2009) and to
Burns and Mitchells (1946) use of 1,277 time series to study business cycles. The
challenge of using large numbers of series is the proliferation of parameters in
standard time series models. While there were large macroeconomic models devel-
oped in the 1960s, and versions of them remain in use today, the restrictions that
reduced the number of parameters in those models were heavily criticized as being
arbitrary, having neither statistical nor economic foundations. Although low-dimen-
sional vector autoregressions had become a standard macroeconometric tool by
the mid-1990s, an outstanding challenge was increasing the number of variables,
both to improve forecasting and to span a wider range of forecast errors, and thus
structural shocks. The technical challenge was that in an unrestricted vector autore-
gression, the number of parameters increases with the square of the number of
variables. Methods were needed to manage this proliferation of parameters if time
series methods were to be used with large numbers of variables.
Dynamic factor models impose parametric restrictions in a way that is consis-
tent with empirical evidence and a broad set of modern theoretical models. In a
dynamic factor model, a given observable variablesay, the growth rate of consump-
tion of servicesis written as the sum of a common component and an idiosyncratic
component. The common component depends on unobserved (or latent) common
variables, called factors, which evolve over time; the idiosyncratic component is
uncorrelated with the common component and has limited correlation with the
other idiosyncratic components. The idiosyncratic component captures measure-
ment error and series-specific disturbances that have no broader macroeconomic
consequences. Thus, in a dynamic factor model, a small number of unobserved
factors explain the comovements of a large number of macroeconomic variables.
This brings us to our next figure, which is taken from Stock and Watson
(2012a). Figure 3 shows the predicted value of six US quarterly macro variables from
a 200-variable, six-factor dynamic factor model; this predicted value is called the
common component of the series. The in-sample R 2 of the common component
for four-quarter growth in GDP (that is, the R 2 of the regression of the four-quarter
growth in GDP on the four-quarter growth of the six factors) is 73 percent; the
average R 2 of the common component over 21 major expenditures variables from
the national income and product accounts is 56 percent; and the average R 2 for all
200 variables is 46 percent. The parameters in this dynamic factor model were fitted
using data from 19592007, so the post-2007 values of the common component
represent the pseudo out-of-sample fit. At the visual level, for these and many other
series, the fit is essentially the same in-sample and out-of-sample, suggesting that the
parameters of the dynamic factor model remained largely stable during and after
the financial crisis.
As Figure 3 illustrates, dynamic factor models fit the data. Techniques for dynamic
factor analysis now can handle arbitrarily many series. One convenient way to estimate
68 Journal of Economic Perspectives

Figure 3
Selected US macroeconomic Time Series: Actual Values and Common Components
(where the common components are the fitted values using the factors from a 200-variable,
6-factor dynamic factor model fit using data from 19592007)

Actual
Common component

Source: Stock-Watson (2012a), Figure 2.

Note: Figure 3 shows the predicted value of six US quarterly macro variables from a 200-variable, 6-factor
dynamic factor model; this predicted value is called the common component of the series. The
parameters in this dynamic factor model were fitted using data from 19592007, so the post-2007 values
of the common component represent the pseudo out-of-sample fit.

the factors is principal components analysis, in which the factors are estimated by least
squares. When estimated using many series, the principal component factor estimates
can be treated as data for subsequent regressions (Stock and Watson 2002; Bai 2003;
Bai and Ng 2006). To implement this approach, one needs to decide how many factors
to use, and Bai and Ng (2002) show how to use information criteria to estimate the
number of factors. This approach can be expanded to arbitrarily many series without
substantially increasing the computational burden, indeed these models provide a
twist on the usual curse of dimensionality: in dynamic factor models, the precision
 James H. Stock and Mark W. Watson 69

of the estimation of the factors improves as the number of data series increases, so
that the curse becomes a blessing.
Because of theoretical and empirical work over the past 20 years, dynamic factor
models have become a leading method for the joint modeling of large numbers
hundredsof economic time series. Dynamic factor models have natural applications
to macroeconomic monitoring and forecasting, a topic we take up below. They also
can be used to estimate the effect of a structural shock, such as a monetary policy
shock, on multiple economic variables. These economy-wide shocks drive the
common factors, and because the factors can be estimated, the economic shocks can
be estimated up to a nonsingular linear transformation. As a result, the techniques
for shock analysis developed for structural vector autoregressions, including the new
methods discussed above, carry over directly to dynamic factor models. By using many
variables, dynamic factor models can more plausibly capture macro-structural shocks
than can low-dimensional vector autoregressions. Moreover, the estimated structural
impulse response functions are internally consistent across all the variables. In Stock
and Watson (2016), we survey dynamic factor models, with a focus on structural shock
analysis.1
Dynamic factor models are not the only method available for high-dimensional
modeling. A different approach is to use a Bayesian prior distribution over the vector
autoregression parameters to reduce the influence of the data on any one parameter
estimate and thus to reduce the amount of noise across parameter estimates. In some
applications, large numbers of restrictions arise naturally: for example, global vector
autoregression reduces the dimensionality of the vector autoreregression parameter
space by restricting domestic variables to depend on foreign variables only through
a small number of weighted averages of global variables (Chudik and Pesaran 2016).
While this discussion has focused on the development of econometric methods
for analyzing high-dimensional time series models, the other major development
that has facilitated this work is the ready availability of data. The Federal Reserve
Bank of St. Louiss FRED database, which migrated to an online platform in 1995,
has been a boon to researchers and to the general public alike. A recent useful
addition to FRED is FRED-MD, a monthly dataset currently comprised of 128
major economic time series for use in high-dimensional macroeconomic modeling
(McCracken and Ng 2016); a beta-version with quarterly data (FRED-QD) is now
available too. These datasets provide a common testbed for high-dimensional
time series modeling and relieve researchers from the arduous task of updating
a large dataset in response to new and revised data. A more specialized database,
maintained by the Federal Reserve Bank of Philadelphia, archives and organizes

1
A variant of a dynamic factor model is the factor-augmented vector autoregression (Bernanke, Boivin,
and Eliasz 2005), in which one or more of the factors are modeled as observed. For example, because the
Federal Reserve controls the federal funds interest rate, Bernanke, Boivin, and Eliasz (2005) argue that
the target interest rate is itself a macroeconomic factor. Alternatively, factor-augmented vector autore-
gression can be interpreted as augmenting a low-dimensional vector autoregression with information
from a first-step dynamic factor model. See Stock and Watson (2016) for a discussion of the relation
between dynamic factor models and factor-augmented vector autoregressions.
70 Journal of Economic Perspectives

real-time economic data; these data are especially valuable to those who want to
test tools for real-time monitoring and forecasting.

Macroeconomic Monitoring and Forecasting

Two important related functions of macroeconomists in business and govern-

ment are tracking the state of the economy and predicting where the economy is
headed. During the 1960s and 1970s, these two functionsmacroeconomic moni-
toring and macroeconomic forecastingrelied heavily on expert judgment. The
1980s and 1990s saw new efforts by time series econometricians to place macro-
economic monitoring and forecasting on a more scientific footing: that is, to be
replicable, to use methods that are transparent and have well-understood prop-
erties, to quantify uncertainty, and to evaluate performance using out-of-sample
experience. While these advances provided macroeconomic monitoring and
forecasting with a solid foundation, much work remained to be done. This work
included improving methods for quantifying and conveying forecast uncertainty;
dramatically expanding the number of data series that could be used, both to enable
real-time monitoring to use the most recently released information and to improve
forecasts; and developing reliable forecasting tools that take into account the evolu-
tion of the economy. Here, we discuss the first two of these: forecast uncertainty and
macroeconomic monitoring. Issues of model instability go far beyond macroeco-
nomic monitoring and forecasting, so we defer that discussion to the next section.

Estimating and Conveying Forecast Uncertainty

A fundamental problem of economic forecasting is that many economic
variables are inherently very difficult to forecast, and despite advances in data avail-
ability, theory, and computational power, we have not seen dramatic improvements
in forecast accuracy over the past decades. One implication of this observation is
that economic forecasters should focus on communicating not just point estimates,
but likely future ranges or distributions of the variable.
Our next figure highlights the development and adoption of density forecasts
over the past 20 years. Figure 4 is a real-time release of a so-called fan chart from the
Bank of Norways Monetary Policy Report for December 2016. A fan chart commu-
nicates uncertainty by providing a density that describes the distribution of possible
future values of the series being forecast, in this case Norwegian consumer price
inflation.2 The Bank of England was an early leader in the use of density forecasts
and fan charts to communicate uncertainty to the public, and these methods are
now widely adopted. Methods for constructing density forecasts are reviewed in

2
The forecast uncertainty is better communicated in color! See the real thing at the websites of the Bank of
England Inflation Report (http://www.bankofengland.co.uk/publications/Pages/inflationreport) and
the Norges Bank Monetary Policy Report (http://www.norges-bank.no/en/Published/Publications/
Monetary-Policy-Report-with-financial-stability-assessment/).
 Twenty Years of Time Series Econometrics in Ten Pictures 71

Figure 4
Fan Chart (Density Forecast) for Consumer Price Index (CPI) inflation in Norway
(percent; four-quarter change)

6 6

30% 50% 70% 90%

5 5

4 4

3 3

2 2

1 1

0 0

1 1
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019

Source: Reproduced from Chart 2.2c of Norges Bank Monetary Policy Report for December 2016, which
used data from Statistics Norway and Norges Bank.
Notes: This chart shows the distribution of possible future values of Norwegian consumer price inflation,
projections for 2016 Q4 through 2019 Q4.

Elliott and Timmermann (2016, ch. 13), and Corradi and Swanson (2006) survey
methods for evaluating the accuracy of density forecasts.
Beyond the clear communication of uncertainty, the past 20 years of academic
work on forecasting has focused on extending the scientific foundations for
forecasting. These include methods for evaluating forecasts (including density fore-
casts), selecting variables for forecasting, and detecting forecast breakdown. While
judgment will inevitably play a role in interpreting model-based forecasts, a central
goal of this research program is to reduce the amount of judgment involved in
constructing a forecast by developing reliable models and tools for evaluating those
models. For a graduate textbook treatment, see Elliott and Timmerman (2016),
and for additional detail see Elliott and Timmermann (2013).

Macroeconomic Monitoring
Twenty years ago, economists who monitored the economy in real time used
indexes of economic indicators and regression models for updating expectations
of individual releases (such as the monthly employment report), combined with
a large dose of judgment based on a narrative of where the economy was headed.
While this approach uses data, it is not scientific in the sense of being replicable,
using well-understood methods, quantifying uncertainty, or being amenable to later
evaluation. Moreover, this method runs the risk of putting too much weight on
the most recent but noisy data releases, putting too little weight on other data,
72 Journal of Economic Perspectives

and being internally inconsistent because each series is handled separately. Because
knowing the current state of the economy in real-time is an ongoing, arguably
increasingly important responsibility of policymakers, time series econometricians
at central banks and in academia have put considerable effort into improving the
foundations and reliability of real-time macroeconomic monitoring.
Our next figure illustrates a central line of research in macroeconomic moni-
toring: the use of large models, in particular dynamic factor models, to incorporate
real-time data releases to provide an internally consistent framework for estimating
current economic conditions. Figure 5 is taken from the February 10, 2017, weekly
update published by the New York Federal Reserve Bank. The dynamic factor
model used by the New York Fed incorporates the most recently available data on
36 major economic indicators to provide a weekly estimate of the growth of GDP
in the current quarter. Figure 5 shows the evolution of this real-time forecast of
current-quarter GDP growthfor obvious reasons, called a nowcast of GDPfor
the fourth quarter of 2016.
In August, the prognosis was for growth slightly above 2 percent at an annual
rate, but by the first Friday in the fourth quarter (October 7), the nowcast had fallen
to 1.3 percent. The November 18 nowcast rose to 2.4 percent on the strength of
retail sales and housing starts data released that week. Then weak industrial produc-
tion data, along with weak housing data released less than two hours before the
December 16 update, pushed that nowcast down to 1.8 percent. As it happened,
the advance estimate of fourth-quarter GDP growth released January 27 was
1.9 percent, slightly less than the estimate of 2.1 percent made on January 20.
Under the hood of this real-time tracking product is a powerful set of tools
for updating estimated factors in dynamic factor models using real-time data flows.
The use of dynamic factor models for real-time macroeconomic monitoring incor-
porating staggered data releases dates to the NBER experimental coincident index
(Stock and Watson 1989). By todays standards, that index was primitive: a monthly
release that encompassed only four variables. The current suite of tools for handling
large series and complicated data flows are exposited in detail in Babura, Gian-
none, Modugno, and Reichlin (2013). The New York Feds model is updated (using
the Kalman filter) as new data arrives, yielding an updated estimate of the single
latent factor which in turn provides an updated estimate of the current-quarter
value of GDP growth. By using a single flexible model, the news content of each
series is exploited in a disciplined and internally consistent way. Some announce-
ments contain substantial news, but many do not, and using a single model to
evaluate these releasesrather than a suite of small models or judgmentprovides
a scientific way to use the real-time data flow.
The New York Fed report is one of several that use dynamic factor models
to provide real-time, publicly available reports on the state of the economy. The
EUROCOIN index, maintained by the Centre for Economic Policy Research and
the Bank of Italy, is a real-time monthly index computed using a dynamic factor
model with approximately 145 variables, calibrated to estimate monthly euro-
zone GDP growth (Altissimo, Cristadoro, Forni, Lippi, and Veronese 2010). The
Chicago Fed National Activity Index is a monthly index of real economic activity
 James H. Stock and Mark W. Watson 73

Figure 5
Contributions of Daily Data Releases to the Federal Reserve Bank of New York Real-
Time Nowcast of 2016Q4 GDP Growth
(bars represent weekly contributions of data revisions to changes in the nowcast)

3.5 3.5

3.0 3.0

2.5 2.5

Percent (annual rate)

Percent (annual rate)

2.0 2.0

1.5 1.5

1.0 1.0

0.5 0.5

0 0

0.5 0.5

1.0 1.0
Au

Ja
O
Se

Se

Se

Fe
N

Ja
D

n
ct
g

ct

ov

ov

ec

ec

n
p

b
19

06
30

20
02

16

30

14

03
11

25

09

23
,2

,2
01

01
6

Advance GDP estimate Retail and consumption

FRBNY Staff Nowcast Income

Housing and construction Labor

Manufacturing International trade

Surveys Others

Source: Federal Reserve Bank of New York Nowcasting report, February 10, 2017.
Note: Figure 5 shows the evolution of a real-time forecast of 2016 fourth-quarter GDPfor obvious
reasons, sometimes called a nowcast. Technically, the points through September 31, 2016, are forecasts
of fourth quarter GDP growth; the points October 1 through December 31, 2016, are nowcasts; and the
points January 1, 2017, to the end of the series are backcasts of fourth quarter GDP growth.

constructed as the single factor in an 85-variable dynamic factor model. The Federal
Reserve Bank of Philadelphia maintains the Aruoba-Diebold-Scotti (2009) index,
which is updated weekly using a six-variable dynamic factor model with one quar-
terly series (GDP), four monthly series, and one weekly series. The Federal Reserve
Bank of Atlantas real-time nowcasting tool, GDPNow, uses a dynamic factor model
combined with a GDP accounting approach to estimate current-quarter GDP.
There are other methods for nowcasting and mixed-frequency data. One
popular tool for single-equation prediction using mixed-frequency data is the
MIDAS model (Ghysels, Sinko, and Valkanov 2007), in which high-frequency data
74 Journal of Economic Perspectives

are temporally aggregated using data-dependent weights. For a survey of methods

of mixed-frequency nowcasting and forecasting, see Foroni and Marcellino (2013).

Model Instability and Latent Variables

A large empirical literature has documented instability in both large- and small-
dimensional time series models. A particularly well-known example of this instability
is the Great Moderation, the period from 1984 to 2007 in which the volatility of many
macroeconomic time series was greatly reduced. Examples of some of the many papers
that document instability in the parameters of time series models include Stock and
Watson (1996) for univariate time series forecasts, Stock and Watson (2003) for infla-
tion forecasts using asset prices, and Welch and Goyal (2008) for equity premium
forecasts. The methods in this literature draw in part on tests for breaks, time varia-
tion, and out-of-sample stability that date to the early 1990s.
This widespread nature of instability in time series relations raises the question
of how to modify time series models so that they can be useful even in the presence
of instability. An early approach was to model instability as deterministic regime
shifts, but while useful, that approach is often unsatisfying because, outside of appli-
cations to a policy regime shift, the single-break model is an approximation and in
any event there is rarely a reason to think that another shift will not occur. After
all, the Great Moderation was followed by the financial crisis. A more appealing
modeling strategy is to allow model parameters to evolve over time according to
a stochastic process. If those time-varying parameters multiply observed variables,
then the model has a linear state space (hidden Markov) structure and the Gaussian
likelihood can be computed (using the Kalman filter). If, however, the time-varying
parameters multiply latent variables, then it has an inherently nonlinear structure.
Estimating such models is challenging, and it was clear 20 years ago that the rudi-
mentary methods available needed to be improved.
The next two figures illustrate developments in the estimation of nonlinear latent
variable models over the past 20 years. The first, Figure 6, is from Kim and Nelson
(1999); the figure shows real GDP growth (the solid line), and the posterior prob-
ability of a break in the variance in GDP (dashed line). Based on this figure, Kim and
Nelson (1999) concluded that US GDP growth had entered a period of low volatility,
and that the most likely date for this transition was 1984Q1. This conclusion was reached
independently using break test methods by McConnell and Perez-Quiros (2000). This
low-volatility period, which lasted through 2007 (and to which the economy seems to
have returned) subsequently became known as the Great Moderation.
Aside from its seminal empirical finding, Figure 6 illustrates a major meth-
odological development in handling nonlinear and/or non-Gaussian time series
models with latent variables. Kim and Nelsons (1999) model falls in this category: it
allows for a one-time shift in the mean and variance of GDP growth, layered on top
of Hamiltons (1989) stochastic regime shift model with recurrent shifts in the mean
(which, Hamilton found, aligned with business cycle turning points). A challenge
in these models is estimating the time path of the latent variable given all the data,
 Twenty Years of Time Series Econometrics in Ten Pictures 75

Figure 6
US GDP Growth and the Posterior Probability of a Regime Change in its Innovation
Variance

Source: Kim and Nelson (1999), Fig. 3.A.

Note: The figure shows real GDP growth (the solid line), and the posterior probability of a break in the
variance in GDP (dashed line). Based on this figure, Kim and Nelson (1999) concluded that US GDP
growth had entered a period of low volatility, and that the most likely date for this transition was 1984Q1.

the so-called smoothing problem, along with the model parameters. To estimate
their parameters and to solve this smoothing problemto produce Figure 6Kim
and Nelson used Markov Chain Monte Carlo methods, which break down their
complicated nonlinear non-Gaussian model into a sequence of Monte Carlo simu-
lations using simpler models. Over the past 20 years, Markov Chain Monte Carlo
has become a widely used tool for estimating seemingly intractable nonlinear/non-
Gaussian models. With this tool, Kim and Nelson were able to obtain the posterior
distribution of a one-time structural break in the variance which, as Figure 6 shows,
strongly points to a reduction in the variance of GDP growth early in 1984.
The next figure, Figure 7, shows two panels from Cogley and Sargent (2015) that
illustrate the incorporation of stochastic volatility into latent state variables. Cogley
and Sargent use a univariate model that decomposes the rate of inflation into unob-
served permanent and transitory (measurement error) components, both of which
have innovations with time-varying variances. These variances are modeled as latent
stochastic volatility processes. From a technical perspective, the situation is similar
to that faced by Kim and Nelson (1999) in that the resulting model expresses the
observed data as a nonlinear function of unobserved random variables (the perma-
nent and transitory components of inflation and their volatilities). While the details
differ, the CogleySargent model is also readily estimated by Markov Chain Monte
Carlo methods.
76 Journal of Economic Perspectives

Figure 7
Trend Inflation (Upper Panel) and the Standard Deviation of the Trend
Innovation (Lower Panel) in an Unobserved ComponentsStochastic Volatility
Model of US Inflation, 18502012

Median
0.08 Interquartile range

0.06

0.04

0.02

0.02

1850 1900 1950 2000

q1/2
t

0.015

0.01

0.005

0
1850 1900 1950 2000

Source: Cogley-Sargent (2015), Fig. 7(A, C).

Note: Figure 7 illustrates the incorporation of stochastic volatility into latent state variables. Cogley and
Sargent (2015) use an unobserved-components/stochastic-volatility model to study the evolution of the
US inflation process from 1850 to 2012. Their posterior estimate of trend inflation is shown in the first
panel, and their estimate of the time-varying standard deviation of changes in the trend is shown in the
second panel. They find the periods of greatest variance in the trend to be during the Civil War and
during the period of inflation and disinflation in the 1970s and early 1980s.
 James H. Stock and Mark W. Watson 77

Cogley and Sargent (2015) use this unobserved-components/stochastic-volatility

model to study the evolution of the US inflation process from 1850 to 2012. Their
posterior estimate of trend inflation is shown in the first panel, and their estimate of
the time-varying standard deviation of changes in the trend is shown in the second
panel. They find the periods of greatest variance in the trend to be during the Civil
War and during the period of inflation and disinflation in the 1970s and early 1980s.
The literature on nonlinear/non-Gaussian filtering is complex, nuanced, and
massive. See Durbin and Koopman (2012) for a textbook treatment of linear and
nonlinear filtering methods.

More Reliable Inference

Finally, the past 20 years has seen important work that aims to improve the
quality of statistical inferences. In the mid-1990s, several influential studies found
that widely used methods for computing test statistics with time series data could
reject far too often or, said differently, that confidence intervals could fail to include
the true parameter value far less frequently than the claimed 95 percent coverage
rate. Theoretical econometricians recognized that more work was needed, particu-
larly in the areas of instrumental variables where the instrument might be weak,
standard errors for regression with serially correlated errors, and regression with
highly persistent regressors.

Weak Instruments and Weak Identification

A weak instrument has a small correlation with the variable it is instrumenting,
given the other included variables. For decades, conventional wisdom held that a
weak instrument would simply produce large standard errors, which would correctly
convey that the information in that variable is scant. But a series of papers in the
1990s showed that the consequences of a so-called weak instrument were more
serious: the estimator will in general be biased, conventional standard errors are
misleading, and these problems can occur in very large samples.3 This problem,
which is more generally referred to as weak identification, also arises in generalized
method of moments estimation. Although weak instruments have received the most
attention in microeconometrics, the inferential challenges posed by weak identi-
fication also have played a role in time series econometrics over the past 20 years.
The next figure, taken from Mavroeidis, Plagborg-Mller, and Stock (2014),
illustrates the problems with using conventional asymptotic standard errors and
confidence intervals in instrumental variables methods when one has weak instru-
ments. Figure 8 shows confidence sets for two key parameters of the hybrid New
Keynesian Phillips Curve; on the vertical axis, is the coefficient on marginal cost
(or, in other specifications, the unemployment gap or output gap) and, on the

3
Key papers on this subject from the 1990s include Nelson and Startz (1990a, 1990b) and Hansen,
Heaton, and Yaron (1996) (Monte Carlo simulations), Bound, Jaeger, and Baker (1995) (empirical
application), and Staiger and Stock (1997) (econometric theory).
78 Journal of Economic Perspectives

Figure 8
Point Estimate and 90% Confidence Sets for Hybrid New Keynesian Phillips
Curve Parameters: Standard Generalized Method of Moments (Ellipse) and Weak-
Instrument Robust (Gray)

Source: Mavroeidis, Plagborg-Mller, and Stock (2014), Fig. 11a.

Note: Figure 8 shows confidence sets for two key parameters of the hybrid New Keynesian Phillips Curve.
The dot is the point estimate using generalized method of moments, and the small ellipse around
the point estimate is the corresponding nominal 90 percent confidence set computed using textbook
asymptotics. The gray regions in the figure comprise a 90 percent confidence set that is robust to the
use of weak instruments. The figures show that the weak-identification robust confidence sets differ
dramatically from the standard asymptotic confidence ellipse. See text for details.

horizontal axis, f is the coefficient on forward-looking rational expectations (some-

times interpreted as relating to the fraction of forward-looking agents). The results
in this figure were computed using data from 19842011, where, following Gal and
Gertler (1999), the labor share is the proxy for marginal cost, and the instruments
are three lags each of marginal cost and the change in inflation, pruned down from
Gal and Gertlers (1999) original set of 24 instruments (which yield similar qualita-
tive results). The dot is the point estimate using generalized method of moments,
and the small ellipse around the point estimate is the corresponding nominal 90
percent confidence set computed using textbook asymptotics. The gray regions in
the figure comprise a 90 percent confidence set that is robust to the use of weak
instruments. The obvious conclusion from Figure 8 is that the weak-identification
robust confidence sets differ dramatically from the standard asymptotic confi-
dence ellipse. Mavroeidis, Plagborg-Mller, and Stock (2014) argue that the reason
for this divergence is that the instruments used in this generalized method of
moments estimation are weak. This problem of weak identification arises broadly
in New Keynesian Phillips Curve applications (for example, Henry and Pagan 2004;
Mavroeidis 2004; Nason and Smith 2008).
Weak identification also arises in other contexts, like in the estimation of inter-
temporal consumption-based asset pricing models (Stock and Wright 2000) and
estimation of monetary policy reaction functions using generalized method of
moments (Consolo and Favero 2009). Weak identification arises in some types of
 Twenty Years of Time Series Econometrics in Ten Pictures 79

inference in structural autoregressions (for example, Pagan and Robertson 1998;

Chevillon, Mavroeidis, and Zhan 2016; for more references, see Stock and Watson
2016, Section 4). It also arises in complicated ways in the estimation of dynamic
stochastic equilibrium models (for example, Andrews and Mikusheva 2015; Qu 2014).
In linear instrumental variable regressions, one commonly used diagnostic
is to check if the F -statistic testing the hypothesis that the coefficient(s) on the
instrument(s) in the first stage of two stage least squaresthe so-called first-stage
F -statisticis less than 10; if so, weak identification is potentially a problem. This
specific approach is specialized to the homoskedastic setting with uncorrelated
errors; approaches to extending this to heteroskedasticity are proposed by Montiel
Olea and Pflueger (2013) and Andrews (2016).
In the simplest modelsthe textbook regression model with a single
endogenous regressor and errors that are homoskedastic and serially uncorre-
latedthere are now methods for dealing with weak instruments with very good
size and power, both asymptotically and in finite samples. As one departs from this
model, most notably when the number of parameters gets large and/or the model
is nonlinear in the parameters, the toolkit is less complete and theoretical work
remains under way.

Inference with Serially Correlated and Potentially Heteroskedastic Errors

In time series data with a serially correlated error term, each additional obser-
vation does not provide entirely new information about the regression coefficient.
Moreover, many time series regressions exhibit clear signs of heteroskedasticity. In
this setting, the ordinary least squares standard error formula does not apply and
instead standard errors that are robust to heteroskedasticity and autocorrelation
must be used. For example, this problem arises when the dependent variable is a
multi-period return or a multiple-period-ahead variable. The problems of hetero-
skedasticity and autocorrelation also arise in generalized method of moments
models when the data are serially correlated.
In practice, the most commonly used standard errors that are heteroskedas-
ticity- and autocorrelation-robust are computed using methods from seminal papers
by Newey and West (1987) and Andrews (1991). These methods compute standard
errors by replacing the estimate of the variance of the product of the regressor and
the error in the usual heteroskedasticity-robust formula for the variance of the ordi-
nary least squares estimator with a weighted average of the autocovariances of that
product; the number of autocovariances averaged is determined by the so-called
bandwidth parameter. But even 20 years ago, there were inklings that the perfor-
mance of hypothesis tests and confidence intervals constructed using these standard
errors in typical macroeconometric applications fell short of the asymptotic perfor-
mance used to justify the tests. In an early Monte Carlo simulation, den Haan and
Levin (1997) studied the rejection rates of tests using these standard errors under
the null hypothesisthat is, the size of the test. Depending on the persistence in
the data, they found that a test that should reject 5 percent of the time under the
null will in practice reject 10 or even 20 percent of the time. If the aim of a research
project is, say, to test for predictability in multiyear stock returns using monthly
80 Journal of Economic Perspectives

data, this over-rejection could easily lead to an incorrect conclusion that returns are
predictable when in fact they are not.
Understanding the source of these size distortions and improving upon
NeweyWest/Andrews standard errors therefore became a major line of research
by theoretical econometricians over the past 20 years, which is succinctly surveyed
by Mller (2014, Sections 23). In brief, this line of work finds that to construct
tests with a rejection rate closer to the desired 5 percent, it is necessary to use band-
widths much larger than those suggested by NeweyWest and Andrews. But doing
so results in a complication: the test statistic no longer has the usual large-sample
normal distribution and, in general, nonstandard critical values must be used.
These ideas were set out by Kiefer, Vogelsang, and Bunzel (2000), and their insights
prompted a large literature aimed at understanding and refining their large-band-
width approach. This theoretical literature has now produced multiple methods
that yield far smaller size distortions than tests based on NeweyWest/Andrews stan-
dard errors, and which also have better power than the KieferVogelsangBunzel
test. Moreover, some of these tests have standard critical values, simplifying their
use in practice.
Applied econometricians typically are eager to use the most recent econometric
method when they demonstrably improve upon the methods of the past. Curiously,
this has not been the case for heteroskedasticity- and autocorrelation-robust infer-
ence, where empirical practice continues to be dominated by NeweyWest/Andrews
standard errors. The new methods are easy to use, straightforward to understand,
and have a lineage that traces back 40 years. It is time for empirical researchers in
time series econometrics to take the next step and to adopt these improved methods
for heteroskedasticity- and autocorrelation-robust inference.

Long-run Relations, Cointegration, and Persistent Regressors

The basic insight of cointegrationthe development for which Clive Granger
received the Nobel Prize in 2003is that multiple persistent macroeconomic vari-
ables move together at low frequencies, that is, they share common long-term trends.
Moreover, these low-frequency comovements connect with basic economic theories
such as balanced economic growth. But while there was a surge of work on cointegra-
tion in the 1980s and 1990s, such work has received less emphasis since then.
Our final historical figure, from Elliott (1998), illustrates a technical road-
block hit by this research program. Elliotts figure, our Figure 9, portrays the null
rejection rate of a test of the value of a cointegrating coefficient in a simple model
with two cointegrated variables. The test maintains that each of the variables is
integrated of order one, that is, has a unit autoregressive root, an assumption that
is part of the cointegration model. Figure 9 shows that small departures from this
unit-root assumption (as measured by c, which is the difference between the true
largest root and one, multiplied by the sample size) can cause major problems for
tests and confidence intervals about the value of that cointegrating coefficient:
tests that are supposed to reject 5 percent of the time under the null can reject
with very high rates (shown on the vertical axis), particularly when the correla-
tion (shown on the horizontal axis) between innovations in the error and in
 James H. Stock and Mark W. Watson 81

Figure 9
Asymptotic Size of Tests of Values of the Cointegrating Coefficient Using Efficient
Cointegrating Estimators and Their Standard Errors when the Time Series Follow
Local-to-Unity Processes with Parameter c
(delta is the correlation between innovations in the error and in the regressor)

Source: Elliott (1998), Figure 1(a).

Note: The figures portrays the null rejection rate of a test of the value of a cointegrating coefficient in a
simple model with two cointegrated variables. The test maintains that each of the variables is integrated
of order one, that is, has a unit autoregressive root, an assumption that is part of the cointegration model.
The figure shows that small departures from this unit-root assumption (as measured by c, which is the
difference between the true largest root and one, multiplied by the sample size) can cause major problems
for tests and confidence intervals about the value of that cointegrating coefficient.

the regressor is large. In fact, this problem arises for deviations from a unit root
that are too small to be detected with high probability, even in arbitrarily large
samples. As a result, standard methods of inference developed for cointegration
models are not robust to effectively undetectable departures from the model,
making such inference unreliable.
While subsequent work has produced novel ideas by econometric theorists,
the proposed methods have drawbacks and no alternative set of procedures have
emerged. In fact, the literature has shown that the problem documented in Figure
9 goes beyond the local-to-unity model used by Elliott (1998) and other researchers
in this area. Related problems of inference also arise in regressions in which a
regressor is persistent, as can occur in applications with financial data.
It is important to stress that these challenges are technical ones; the basic insight
of cointegration that variables move together at low frequencies is a deep one that
connects with core economic theories such as balanced growth and the term struc-
ture of interest rates. But inference, and perhaps modeling, of those comovements
can be more complicated than had originally been thought.
82 Journal of Economic Perspectives

Figure 10
The Mother of All Forecast Errors: Survey of Professional Forecasters Median
Forecast for Nonfarm Business Employment during the 20072009 Recession and
Early Recovery

142,000

140,000

138,000
Thousands of jobs

136,000

134,000

132,000

130,000

128,000
2005 2006 2007 2008 2009 2010 2011

Source: Philadelphia Fed Survey of Professional Forecasters.

Note: This figure shows the real-time median forecast of the log of nonfarm employment recorded by the
Survey of Professional Forecasters in the quarters leading up to and through the financial crisis. Even
well after the crisis began and real-time information about the collapse of the economy was available,
these forecasters consistently predicted a mild recession.

Challenges Ahead

We close by mentioning a few of the research challenges for time series

econometrics. Our final figure shows that despite the substantial improvements in
forecasting methods over the past decades, much work remains. When we teach,
we call Figure 10 the Mother of All Forecast Errors. This figure shows the real-
time median forecast of the log of nonfarm employment recorded by the Survey
of Professional Forecasters in the quarters leading up to and through the financial
crisis. Even well after the crisis began and real-time information about the collapse
of the economy was available, these forecasters consistently predicted a mild reces-
sion. A small part of these errors is due to revisions between preliminary and final
data, but most of these errors, we believe, represent a failure of forecasting models
to capture the severity of the shocks and their devastating effect on the economy.
Forecasters certainly were not the only economists to misjudge events leading up to
and during the financial crisis! But this is an article about time series methods, and
in our view, tackling the challenge of Figure10 is a priority.
Another open challenge lies in the big data sphere. The methods of the past
20 yearsdynamic factor models and large Bayesian vector autoregressionshave
made it possible to include arbitrarily many series in forecasting systems and to incor-
porate data releases in real time, and the result has been large improvements in
 Twenty Years of Time Series Econometrics in Ten Pictures 83

macroeconomic monitoring. However, there is some evidence that the parametric

restrictions (or priors) that make these methods work discard potentially important
information. In the context of dynamic factor models, the question is whether there
is useful information in the higher factors beyond the handful that would normally be
included (such as the six factors used to produce Figure 3). Some studies have looked
at this question, with mixed results; for example, Carrasco and Rossi (2016) give some
positive results, while we give some negative results in Stock and Watson (2012b). A
more ambitious question is whether there is exploitable nonlinear structure in these
data that could perhaps be revealed by modern machine learning methods. While it
is tempting to dive in and use a battery of machine learning methods to attack these
data, one must remember that data snooping can lead to unintentional overstate-
ment of results. One advantage of dynamic factor models, after all, is that they are
closely linked to dynamic macro models (Sargent 1989; Boivin and Giannoni 2006).
We suspect that the next steps towards exploiting additional information in large
datasets will need to use new statistical methods guided by economic theory.
Separately, there are important open questions relating to low-frequency time
series econometrics. For example, what does historical evidence tell us about whether
the recent slowdown in US productivity is permanent or temporary? The answer to
this question is crucial for many long-term economic issues, such as the future of
Social Security and valuing policies to mitigate climate change. Another, technically
related set of questions returns to the basic insight of cointegration and the challenge
posed by Elliotts (1989) figure (Figure 9): there are clearly low-frequency comove-
ments in the data, and macroeconometricians need a set of tools for quantifying those
comovements that does not hinge on adopting a particular model, such as a unit
root model, for the underlying trends. These are technically difficult problems, and
Mller and Watson (2016a, 2016b) propose possible avenues for tackling them.
Finally, there are a number of opportunities for expanding identification and
estimation of macro models by using information in microeconometric data. Here,
opportunities range from estimation of parameters describing individual prefer-
ences and firm behavior, to the possibility of using rich micro data to improve macro
monitoring and forecasting.
The earliest empirical work in macroeconomics relied on time series data;
indeed the first instrumental variables regression was estimated in 1926 using time
series data. The past 20 years has seen a continuation of the vigorous development
of methods for using time series data. These methods draw on improved computa-
tional capacity, better data availability, and new understandings in econometric and
statistical theory. The core driver of these developments is the need of policymakers
for reliable guidance on the effects of contemplated policies, along with their
shared need with the private sector to understand where the economy is and where
it is going. Those needs will not go away. If anything, they become more urgent in
our volatile and ever-changing economic environment. Although the challenges
facing time series econometricians are difficult, so have they been in the past, and
exciting and highly relevant research programs beckon.

We thank Gray Calhoun, Mikkel Plagborg-Mller, and the authors of the papers from which
we took figures, for helpful comments.
84 Journal of Economic Perspectives
References
At-Sahalia, Yacine, and Lars Peter Hansen.
2010. Handbook of Financial Economics, Vol. 1: Tools
and Techniques and Vol. 2: Applications. Elsevier.
Altissimo, Filippo, Riccardo Cristadoro, Mario
Forni, Marco Lippi, and Giovanni Veronese. 2010.
New EuroCOIN: Tracking Economic Growth in
Real Time. Review of Economics and Statistics 92(4):
102434.
Andrews, Donald W. K. 1991. Heteroskedas-
ticity and Autocorrelation Consistent Covariance
Matrix Estimation. Econometrica 59(3): 817858.
Andrews, Isaiah. 2016. Valid Two-Step Identifi-
cation-Robust Confidence Sets for GMM. http://
economics.mit.edu/files/11848.
Andrews, Isaiah, and Anna Mikusheva. 2015.
Maximum Likelihood Inference in Weakly Identi-
fied Dynamic Stochastic General Equilibrium
Models. Quantitative Economics6(1): 12352.
Aruoba, S. Boraan, Francis X. Diebold, and
Chiara Scotti. 2009. Real-Time Measurement
of Business Conditions. Journal of Business &
Economic Statistics 27(4): 41727.
Bai, Jushan. 2003. Inferential Theory for
Factor Models of Large Dimensions. Econometrica
71(1): 13572.
Bai, Jushan, and Serena Ng. 2002. Determining
the Number of Factors in Approximate Factor
Models. Econometrica 70(1): 19121.
Bai, Jushan, and Serena Ng. 2006. Confidence
Intervals for Diffusion Index Forecasts and
Inference for Factor-Augmented Regressions.
Econometrica 74(4): 113350.
Bank of Norway. 2016. Monetary Policy
Report with Financial Stability Assessment
4/16. December 15, 2016. http://www.norges-
bank.no/en/Published/Publications/
Monetary-Policy-Report-with-financial-stability-
assessment/416-Monetary-Policy-Report/.
Babura, Marta, Domenico Giannone, Michele
Modugno, and Lucrezia Reichlin. 2013. Now-
Casting and the Real-Time Data Flow. Chap. 4 in
Handbook of Economic Forecasting, vol. 2, edited by
Graham Elliott, and Alan Timmermann. Elsevier,
North-Holland.
Baumeister, Christiane, and James D. Hamilton.
2015. Sign Restrictions, Structural Vector
Autoregressions, and Useful Prior Information.
Econometrica 83(5): 196399.
Bernanke, Ben S., Jean Boivin, and Piotr Eliasz.
2005. Measuring the Effects of Monetary Policy: A
Factor-Augmented Vector Autoregressive (FAVAR)
Approach. Quarterly Journal of Economics 120(1):
387422.
Bernanke, Ben S., and Kenneth N. Kuttner.
2005. What Explains the Stock Markets Reaction
to Federal Reserve Policy? Journal of Finance 60(3):
122157.
Boivin, Jean, and Marc Giannoni. 2006. DSGE
Models in a Data-Rich Environment. NBER
Working Paper 12772.
Bound, John, David A. Jaeger, and Regina M.
Baker. 1995. Problems with Instrumental Vari-
ables Estimation When the Correlation between
the Instruments and the Endogenous Explanatory
Variable Is Weak. Journal of the American Statistical
Association 90: 44350.
Burns, Arthur F., and Wesley C. Mitchell. 1946.
Measuring Business Cycles. NBER.
Canova, Fabio. 2007. Methods for Applied Macro-
economic Research. Princeton University Press.
Chevillon, Guillaume, Sophocles Mavroeidis,
and Zhaoguo Zhan. 2016. Robust Inference in
Structural VARs with Long-Run Restrictions.
https://sites.google.com/site/sophoclesmav-
roeidis/research/lrsvar.pdf?attredirects=0.
Chudick, Alexander, and M. Hashem Pesaran.
2016. Theory and Practice of GVAR Modelling.
Journal of Economic Surveys 30(1): 16597.
Cochrane, John H., and Monica Piazzesi. 2002.
The Fed and Interest Rates: A High-Frequency
Identification. American Economic Review 92(2):
9095.
Cogley, Timothy, and Thomas J. Sargent. 2015.
Measuring Price-Level Uncertainty and Insta-
bility in the United States, 18502012. Review of
Economics and Statistics 97(4): 82738.
Consolo, Agostino, and Carlo A. Favero. 2009.
Monetary Policy Inertia: More a Fiction than a
Fact? Journal of Monetary Economics 56(6): 900906.
Corradi, Valentina, and Norman R. Swanson.
2006. Predictive Density Evaluation. Chap. 5
in Handbook of Economic Forecasting, vol. 1, edited
by Graham Elliott, Clive W. J. Granger, and Allan
Timmermann. Elsevier: North Holland.
Corrasco, Marine, and Barbara Rossi. 2016.
In-Sample Inference and Forecasting in Misspeci-
fied Factor Models. Journal of Business and Economic
Statistics 34(3): 31338.
DeJong, David N., Beth F. Ingram, and Charles
H. Whiteman. 2000. A Bayesian Approach to
Dynamic Macroeconomics. Journal of Econometrics
98(2): 203223.
den Haan, W. J., and A. Levin. 1997. A
Practitioners Guide to Robust Covariance Matrix
Estimation. Chap. 12 of Handbook of Statistics, Vol.
15: Robust Inference, edited by G. S. Maddala and C.
R. Rao. Elsevier.
Durbin, J., and S. J. Koopman. 2012. Time Series
Analysis by State Space Methods, 2nd edition. Oxford
University Press.
Elliott, Graham. 1998. On the Robustness of
Cointegration Methods When Regressors Almost
Have Unit Roots. Econometrica 66(1): 14958.
Elliott, Graham, and Alan Timmermann, eds.
2013. Handbook of Economic Forecasting, vol. 2.
Elsevier.
Elliott, Graham, and Alan Timmermann. 2016.
Economic Forecasting. Princeton University Press.

Faust, Jon, John H. Rogers, Eric Swanson,
and Jonathan H. Wright. 2003. Identifying the
Effects of Monetary Policy Shocks on Exchange
Rates Using High Frequency Data. Journal of the
European Economic Association 1(5): 103157.
Federal Reserve Bank of New York. 2017.
Nowcasting Report, February 10, 2017. Available
at: https://www.newyorkfed.org/research/policy/
nowcast.
Fernndez-Villaverde, Jess, Juan F. Rubio-
Ramrez, and Frank Schorfheide. 2016. Solution
and Estimation Methods for DSGE Models. Chap.
9 in Handbook of Macroeconomics, vol. 2, edited by
John B. Taylor and Harald Uhlig. Elsevier.
Foroni, Claudia, and Massimiliano Marcellino.
2013. A Survey of Econometric Methods for
Mixed-Frequency Data. Norges Bank Research
Working Paper 2013-6.
Friedman, Walter F. 2009. The Harvard
Economic Service and the Problems of Fore-
casting. History of Political Economy 41(1): 5788.
Fry, Rene, and Adrian Pagan. 2011. Sign
Restrictions in Structural Vector Autoregressions:
A Critical Review. Journal of Economic Literature
49(4): 93860.
Gal, Jordi, and Mark Gertler. 1999. Inflation
Dynamics: A Structural Econometric Analysis.
Journal of Moneteary Economics 44(2): 195222.
Gertler, Mark, and Peter Karadi. 2015. Mone-
tary Policy Surprises, Credit Costs, and Economic
Activity. American Economic Journal: Macroeconomics
7(1): 4476.
Ghysels, Eric, Arthur Sinko, and Rossen
Valkanov. 2007. MIDAS Regressions: Further
Results and New Directions. Econometric Reviews
26(1): 5390.
Giacomini, Raffaella, and Toru Kitagawa. 2014.
Inference about Non-Identified SVARs. CEPR
Discussion Paper 10287.
Hamilton, James D. 1989. A New Approach
to the Economic Analysis of Nonstationary Time
Series and the Business Cycle. Econometrica 57(2):
35784.
Hamilton, James D. 2003. What Is an Oil
Shock? Journal of Econometrics 113(2): 36398.
Hansen, Lars P., John Heaton, and Amir Yaron.
1996. Finite Sample Properties of Some Alterna-
tive GMM Estimators. Journal of Business and
Economic Statistics 14(3): 26280.
Henry, S. G. B., and A. R. Pagan. 2004. The
Econometrics of the New Keynesian Policy Model:
Introduction. Oxford Bulletin of Economics and
Statistics 66(Supplement): 581607.
Herbst, Edward P., and Frank Schorfheide.
2015. Bayesian Estimation of DSGE Models. Princ-
eton University Press.
Ireland, Peter N. 1997. A Small, Structural,
Quarterly Model for Monetary Policy Evaluation.
Carnegie-Rochester Conference Series on Public Policy
47: 83108.
Kiefer, Nicholas M., Timothy J. Vogelsang,
James H. Stock and Mark W. Watson 85
and Helle Bunzel. 2000. Simple Robust Testing
of Regression Hypotheses. Econometrica 68(3):
695714.
Kilian, Lutz. 2008. Exogenous Oil Supply
Shocks: How Big Are They and How Much Do
They Matter for the U.S. Economy? Review of
Economics and Statistics 90(2): 21640.
Kim, Chang-Jin, and Charles R. Nelson. 1999.
Has the U.S. Economy Become More Stable? A
Bayesian Approach Based on a Markov-Switching
Model of the Business Cycle. Review of Economics
and Statistics 81(4): 608616.
Kuttner, Kenneth N. 2001. Monetary Policy
Surprises and Interest Rates: Evidence from the
Fed Funds Futures Market. Journal of Monetary
Economics 47(3): 52344.
Ltkepohl, Helmut. 2013. Identifying
Structural Vector Autoregressions via Changes in
Volatility. Advances in Econometrics, vol. 32, pp.
169203.
Mavroeidis, Sophocles. 2004. Weak Identifica-
tion of Forward-Looking Models in Monetary
Economics. Oxford Bulletin of Economics and
Statistics 66(Supplement): 60935.
Mavroeidis, Sophocles, Mikkel Plagborg-Mller,
and James H. Stock. 2014. Empirical Evidence
on Inflation Expectations in the New Keynesian
Phillips Curve. Journal of Economic Literature 52(1):
12488.
McConnell, Margret M., and Gabriel Perez-
Quiros. 2000. Output Fluctuations in the United
States: What Has Changed Since the Early 1980s.
American Economic Review 90(5): 146476.
McCracken, Michael W., and Serena Ng. 2016.
FRED-MD: A Monthly Database for Macroeco-
nomic Research. Journal of Business & Economic
Statistics 34(4): 57489.
Mertens, Karel, and Morten O. Ravn. 2013.
The Dynamic Effects of Personal and Corporate
Income Tax Changes in the United States.
American Economic Review 103(4): 121247.
Montiel Olea, Jos Luis, and Carolin Pflueger.
2013. A Robust Test for Weak Instruments.
Journal of Business and Economic Statistics 31(3):
35869.
Moon, Hyungsik Roger, Frank Schorfheide,
and Eleonara Granziera. 2013. Inference for VARs
Identified with Sign Restrictions. http://sites.sas.
upenn.edu/schorf/files/svar-paper.pdf.
Mller, Ulrich K. 2014. HAC Corrections for
Strongly Autocorrelated Time Series. Journal of
Business and Economic Statistics 32(3): 31122.
Mller, Ulrich K., and Mark W. Watson. 2016a.
Measuring Uncertainty about Long-Run Predic-
tions. Review Economic Studies 84(4): 171140.
Mller, Ulrich K., and Mark W. Watson. 2016b.
Long-Run Covariability. Unpublished paper,
Princeton University.
Nason, James M., and Gregor W. Smith. 2008.
Identifying the New Keynesian Phillips Curve.
Journal of Applied Econometrics 23(5): 52551.
86 Journal of Economic Perspectives
Nelson, Charles R., and Richard Startz. 1990a.
The Distribution of the Instrumental Variable
Estimator and Its t Ratio When the Instrument
Is a Poor One. Journal of Business 63(1, Part 2):
S125S140.
Nelson, Charles R., and Richard Startz. 1990b.
Some Further Results on the Exact Small Sample
Properties of the Instrumental Variable Estimator.
Econometrica 58(4): 96776.
Newey, Whitney K., and Kenneth D. West. 1987.
A Simple Positive Semi-Definite, Heteroskedas-
ticity and Autocorrelation Consistent Covariance
Matrix. Econometrica 55(3): 703708.
Normandin, Michel, and Louis Phaneuf. 2004.
Monetary Policy Shocks: Testing Identification
Conditions under Time-Varying Conditional Vola-
tility. Journal of Monetary Economics 51(6): 121743.
Otrok, Christopher. 2001. On Measuring
the Welfare Costs of Business Cycles. Journal of
Monetary Economics 47(1): 6192.
Pagan, A. R., and J. C. Robertson. 1998. Struc-
tural Models of the Liquidity Effect. Review of
Economics and Statistics 80(2): 202217.
Plagborg-Mller, Mikkel. 2016. Bayesian Infer-
ence on Structural Impulse Response Functions.
http://scholar.harvard.edu/files/plagborg/files/
irf_bayes.pdf.
Qu, Zhongjun. 2014. Inference in DSGE
Models with Possible Weak Identification. Quan-
titative Economics 5(2): 45794.
Ramey, V. A. 2016. Macroeconomic Shocks
and their Propagation. Chap. 2 in Handbook of
Macroeconomics vol. 2., edited by J. B. Taylor and H.
Uhlig, 71162. Elsevier.
Rigobon, Roberto. 2003. Identification
through Heteroskedasticity. Review of Economics
and Statistics 85(4): 77792.
Rigobon, Roberto, and Brian Sack. 2003.
Measuring the Reaction of Monetary Policy to
the Stock Market. Quarterly Journal of Economics
118(2): 63969
Rigobon, Roberto, and Brian Sack. 2004. The
Impact of Monetary Policy on Asset Prices. Journal
of Monetary Economics 51(8): 155375.
Romer, Christina D., and David H. Romer.
1989. Does Monetary Policy Matter? A New Test
in the Spirit of Friedman and Schwartz. In NBER
Macroeconomics Annual 1989, vol. 4, edited by
Olivier Blanchard and Stanley Fisher, 12170. MIT
Press.
Rudebusch, Glenn D. 1998. Do Measures of
Monetary Policy in a VAR Make Sense? Interna-
tional Economic Review 39(4): 907931.
Sargent, Thomas J. 1989. Two Models of
Measurements and the Investment Accelerator.
Journal of Political Economy 97(2): 25187.
Schorfheide, Frank. 2000. Loss Function-based
Evaluation of DSGE Models. Journal of Applied
Econometrics 15(6): 64570.
Sentana, Enrique, and Gabriele Fiorentini.
2001. Identification, Estimation, and Testing of
Conditionally Heteroskedastic Factor Models.
Journal of Econometrics 102(2): 14364.
Sims, Christopher A. 1980. Macroeconomics
and Reality. Econometrica 48(1): 148.
Smets, Frank, and Raf Wouters. 2003. An Esti-
mated Dynamic Stochastic General Equilibrium
Model of the Euro Area. Journal of the European
Economic Association 1(5): 112375.
Staiger, Douglas, and James H. Stock. 1997.
Instrumental Variables Regression with Weak
Instruments. Econometrica 65(3): 55786.
Stock, James H. 2008. Whats New in Economet-
rics: Time Series, Lecture 7. Short course lectures,
NBER Summer Institute, at http://www.nber.org/
minicourse_2008.html.
Stock, James H., and Mark W. Watson. 1989.
New Indexes of Coincident and Leading
Economic Indicators. NBER Macroeconomics
Annual 1989, edited by Olivier J. Blanchard and
Stanley Fischer, 35193. MIT Press.
Stock, James H., and Mark W. Watson. 1996.
Evidence on Structural Instability in Macroeco-
nomic Time Series Relations. Journal of Business
and Economic Statistics 14(1): 1130.
Stock, James H., and Mark W. Watson. 2002.
Forecasting Using Principal Components from
a Large Number of Predictors. Journal of the
American Statistical Association 97(460): 116779.
Stock, James H., and Mark W. Watson. 2003.
Forecasting Output and Inflation: The Role of
Asset Prices. Journal of Economic Literature 41(3):
788829.
Stock, James H., and Mark W. Watson. 2011.
Introduction to Econometrics, 3rd Edition. Pearson.
Stock, James H., and Mark W. Watson. 2012a.
Disentangling the Channels of the 200709 Reces-
sion. Brookings Papers on Economic Activity, no. 1,
81135.
Stock, James H., and Mark W. Watson. 2012b.
Generalized Shrinkage Methods for Forecasting
Using Many Predictors. Journal of Business &
Economic Statistics 30(4): 48193.
Stock, James H., and Mark W. Watson. 2016.
Dynamic Factor Models, Factor-Augmented
Autoregressions, and Structural Vector Autoregres-
sions in Macroeconomics. Chap. 8 in Handbook of
Macroeconomics, vol. 2, edited by John B. Taylor and
Harald Uhlig, 415526. Elsevier.
Stock, James H., and Jonathan H. Wright. 2000.
GMM with Weak Identification. Econometrica
68(5): 105596.
Uhlig, Harald. 2005. What are the Effects
of Monetary Policy on Output? Results from an
Agnostic Identification Procedure. Journal of
Monetary Economics 52(2): 381419.
Welch, Ivo, and Amit Goyal. 2008. A Compre-
hensive Look at the Empirical Performance of
Equity Premium Prediction. Review of Financial
Studies 21(4): 14551508.